Extract Text from Multilingual Documents
Extracts text in multiple languages from images with markdown conversion and text grounding. Vision-language model for document layout understanding.
What it does
DeepSeek OCR reads text from images and converts it to searchable, copyable text. Point it at any image with text - a document, sign, receipt - and it automatically finds the text and converts it to digital characters. Supports multiple languages.
Problem it solves
Extract multilingual documents | Automatically detect and extract text in multiple languages from images |
Convert documents to markdown | Preserve document structure (headings, paragraphs, formatting) as readable markdown |
Parse complex layouts | Handle tables, figures, and mixed-format documents |
Locate specific text | Find where specific text appears in an image with bounding box coordinates |
Process diverse document types | Works on scans, PDFs converted to images, photos of documents |
SEO optimization | Generate descriptions for image search |
Product catalogs | Auto-describe products for e-commerce |
Social media | Generate captions for content sharing |
Input/Output
- Input: Any image with text
- Format: JPEG, PNG, BMP, GIF, or PDF
- Resolution: Any size (optimized for up to 2048×2048)
- Content: Photos, scans, documents, screenshots, invoices, papers
- Text type: Printed or handwritten text
- Output: Extracted text with multiple options
Performance
- Accuracy by Text Type:
- Printed text: High accuracy on clear documents
- Handwritten text: Works on legible handwriting
- Clear documents: Excellent on well-formatted text
- Complex layouts: Handles tables and mixed formatting
- Mixed-language images: Detects and extracts multiple languages simultaneously
- Real-World Scenarios:
- Multilingual documents: Excellent for mixed-language processing
- Markdown conversion: Preserves document structure and formatting
- Complex layouts: Handles tables, figures, and complex formatting
- Text grounding: Locates specific text with coordinates
Technical Details
Architecture | DeepSeek V2 (vision-language model) |
Key Components | • Vision Encoder: Processes images into feature representations
• Language Decoder: DeepSeek V2 LLM generates text output
• Attention: MLA (Multi-head Latent Attention) for efficiency |
Inference Pipeline | Vision Encoder processes image → Task-specific prompt → Language Decoder generates output → Returns text, coordinates, or markdown |
Model Variants | • Official: https://huggingface.co/deepseek-ai/DeepSeek-OCR (Transformers 4.46.3)
• Compatible: https://huggingface.co/strangervisionhf/deepseek-ocr-latest-transformers (Transformers 4.57.1+) |
Training | • Framework: PyTorch
• Languages: Multilingual corpus
• Model size: 6.68 GB |
Compliance & Provenance
Provider | Open-source |
Provider type | Specialized |
License | MIT |
EU AI Act risk class | Minimal Risk |
Art. 50 transparency | Not applicable |
Region availability | Available globally |
Training data summary | Pending — provider has not yet published per Art. 53(d) |
For more on how we classify models and mark outputs, see our AI Policy.
Model Source
- GitHub: https://github.com/deepseek-ai/DeepSeek-OCR
- License: MIT License